ECG Foundation Models Show Limited Transfer to Rare Diseases

Foundation models (FMs) trained on vast unlabeled physiological data are a promising avenue for medical AI, but their ability to generalize to rare diseases remains largely unproven. This research specifically examines the transferability of nine publicly available ECG FMs for detecting Brugada syndrome, a rare cardiac condition. The study evaluated these FMs under various training strategies (from-scratch, linear probing, full fine-tuning) and data configurations, including data ablation and zero-shot cross-site transfers. It found that while pre-training was essential for high-capacity architectures to converge, it offered no significant advantage for more compact models that could already converge with labeled data alone. Crucially, the best fine-tuned FM showed only marginal improvement over strong supervised baselines on full datasets, and data-efficiency advantages observed on smaller subsets did not consistently replicate across different cohorts. Furthermore, FM-based pipelines failed to generalize better than supervised baselines in zero-shot cross-site transfers, often performing at chance level. These results suggest that for Brugada syndrome, FM pre-training primarily provides mechanical optimization stability rather than encoding genuinely transferable clinical knowledge, challenging a common assumption about large-scale pre-training.